onnx-array-api: APIs to create ONNX Graphs

onnx-array-api implements APIs to create custom ONNX graphs. The objective is to speed up the implementation of converter libraries. The library is released on pypi/onnx-array-api and its documentation is published at APIs to create ONNX Graphs.

Numpy API

The first one matches numpy API. It gives the user the ability to convert functions written following the numpy API to convert that function into ONNX as well as to execute it.

import numpy as np
from onnx_array_api.npx import absolute, jit_onnx
from onnx_array_api.plotting.text_plot import onnx_simple_text_plot

def l1_loss(x, y):
    return absolute(x - y).sum()


def l2_loss(x, y):
    return ((x - y) ** 2).sum()


def myloss(x, y):
    return l1_loss(x[:, 0], y[:, 0]) + l2_loss(x[:, 1], y[:, 1])


jitted_myloss = jit_onnx(myloss)

x = np.array([[0.1, 0.2], [0.3, 0.4]], dtype=np.float32)
y = np.array([[0.11, 0.22], [0.33, 0.44]], dtype=np.float32)

res = jitted_myloss(x, y)
print(res)

print(onnx_simple_text_plot(jitted_myloss.get_onnx()))

[0.042]
opset: domain='' version=18
input: name='x0' type=dtype('float32') shape=['', '']
input: name='x1' type=dtype('float32') shape=['', '']
Sub(x0, x1) -> r__0
  Abs(r__0) -> r__1
    ReduceSum(r__1, keepdims=0) -> r__2
output: name='r__2' type=dtype('float32') shape=None

It supports eager mode as well:

import numpy as np
from onnx_array_api.npx import absolute, eager_onnx


def l1_loss(x, y):
    err = absolute(x - y).sum()
    print(f"l1_loss={err.numpy()}")
    return err


def l2_loss(x, y):
    err = ((x - y) ** 2).sum()
    print(f"l2_loss={err.numpy()}")
    return err


def myloss(x, y):
    return l1_loss(x[:, 0], y[:, 0]) + l2_loss(x[:, 1], y[:, 1])


eager_myloss = eager_onnx(myloss)

x = np.array([[0.1, 0.2], [0.3, 0.4]], dtype=np.float32)
y = np.array([[0.11, 0.22], [0.33, 0.44]], dtype=np.float32)

res = eager_myloss(x, y)
print(res)

l1_loss=[0.04]
l2_loss=[0.002]
[0.042]

Light API

The second API or Light API tends to do every thing in one line. It is inspired from the Reverse Polish Notation. The euclidean distance looks like the following:

import numpy as np
from onnx_array_api.light_api import start
from onnx_array_api.plotting.text_plot import onnx_simple_text_plot

model = (
    start()
    .vin("X")
    .vin("Y")
    .bring("X", "Y")
    .Sub()
    .rename("dxy")
    .cst(np.array([2], dtype=np.int64), "two")
    .bring("dxy", "two")
    .Pow()
    .ReduceSum()
    .rename("Z")
    .vout()
    .to_onnx()
)    

GraphBuilder API

Almost every converting library (converting a machine learned model to ONNX) is implementing its own graph builder and customizes it for its needs. It handles some frequent tasks such as giving names to intermediate results, loading, saving onnx models. It can be used as well to extend an existing graph.

import numpy as np
from onnx_array_api.graph_api  import GraphBuilder

g = GraphBuilder()
g.make_tensor_input("X", np.float32, (None, None))
g.make_tensor_input("Y", np.float32, (None, None))
r1 = g.make_node("Sub", ["X", "Y"])  # the name given to the output is given by the class,
                                     # it ensures the name is unique
init = g.make_initializer(np.array([2], dtype=np.int64))  # the class automatically
                                                          # converts the array to a tensor
r2 = g.make_node("Pow", [r1, init])
g.make_node("ReduceSum", [r2], outputs=["Z"])  # the output name is given because
                                               # the user wants to choose the name
g.make_tensor_output("Z", np.float32, (None, None))

onx = g.to_onnx()  # final conversion to onnx